MiniMax H3 Ships Open Weights: 33B Omni Video Model Challenges Closed API Leaders at $0.08/s
MiniMax released H3 on July 31 as a fully open-weight model, putting video generation quality that competes with closed API leaders onto consumer and researcher hardware. The weights are on HuggingFace. The model is already running locally on RTX 4090s.
The story here is not just the benchmark position. It is the architecture.
One Model Replacing a Pipeline
Video generation has historically required stacking specialist systems. A text-to-video model. An image animator. A separate audio synthesis step. An upscaler. H3 collapses that pipeline into a single 33B parameter diffusion transformer.
The encoder is Qwen3-VL-32B — a full frontier vision-language model embedded as the perception layer. Inputs are unified: up to nine reference images, three reference videos, and three audio clips can be passed as a single context. The model handles first-frame anchoring, last-frame targets, reference style, and audio grounding in one forward pass.
Output is 4-15 second video at 768P or 2K resolution with 32 kHz native stereo audio — generated alongside the video, not post-processed.
Two Modes
H3 runs in two modes:
FL2VA (First Last Frame to Video/Audio): Takes zero, one, or two images as the first or last frame and interpolates the motion between them. Standard image-to-video generation, but with the audio track generated simultaneously.
Ref2VA (Reference to Video/Audio): Takes up to 12 mixed inputs — images, video clips, audio references — and uses them as style and content anchors. The model treats all inputs as unified context, not a sequential conditioning stack.
Arena Benchmarks
On Artificial Analysis Arena, H3 scores 1193 on Image-to-Video and 1237 on Text-to-Video. SeedDance 2.5, the current leader, holds the top position at approximately 1450 ELO. H3 is not beating the closed leaders — it is compressing the gap to a level no open-weight video model has reached.
On August 13, H3 was added to Arena’s Video Edit leaderboard, completing its presence across the three core video generation categories.
Hardware and Local Inference
The full pipeline that runs through the MiniMax API involves three steps: prompt enhancement, base generation, and 2K upscaling. Only the base generation block is in the open-weight release. Local inference reproduces the base output quality; the API endpoint adds the enhancement and upscaler layers.
Minimum VRAM for local inference: 24GB. With quantization and CPU offloading, the model fits on 8GB. For full-precision weights and longer 15-second generations, 50-60GB is the practical target. A single H100 or H200 runs the full pipeline at inference scale.
Pricing
Through the MiniMax API: $0.08 per second of output at 768P, $0.13 per second at 2K. A 10-second generation costs $0.80 at standard resolution or $1.30 at 2K — substantially cheaper than Runway’s and Sora’s per-generation pricing at comparable quality.
Audio references beyond the base case and additional image references add incremental cost on top of the per-second generation fee. Reference video input is billed at the same per-second rate as output.
Key Numbers
| Metric | Value |
|---|---|
| Parameters | 33B diffusion transformer |
| Encoder | Qwen3-VL-32B |
| Max inputs | 12 (9 images + 3 videos + 3 audio) |
| Output duration | 4-15 seconds |
| Output resolution | 768P or 2K |
| Audio | 32 kHz stereo native |
| Pricing (768P) | $0.08/s generated |
| Pricing (2K) | $0.13/s generated |
| AA I2V Arena ELO | 1193 |
| AA T2V Arena ELO | 1237 |
| Minimum VRAM (local) | 24 GB (8 GB quantized) |
| License | Open weights (HuggingFace) |